Remove inference multiplicity

[carrascomj]

Closes #24

Description

Running inference with --nsamples_per_protein greater than 1 blows up memory using the torch backend. This happens because inference_multiplicity was used to replicate every tensor (including ESM embeddings) nsample_per_protein times before moving them to the GPU, so multi-sample inference pushed huge duplicated feature tensors into vRAM and regularly OOM’d.

After this fix, taking 5 samples of a protein sequence of ~1000 amino acids stays in around 31GiB of vRAM using the torch backend (instead of running OOM for a 81GiB GPU.)

Implementation

I fixed it by removing inference_multiplicity altogether to simply run inference in a nsample_per_protein loop.

This does slow down inference since it's not batched, so feel free to close the PR, just here for future reference.